JumpTableBench

Ground truth for jump table recovery, used to measure the accuracy of B2R2's middle-end on every push.

A jump table is how compilers implement a dense switch: an indirect branch reads a target out of a table in the binary. Recovering those tables is what decides whether a reconstructed CFG is complete, and getting it wrong is silent. This repository pins down the right answer for real binaries per architecture (x86 and x64) so that a regression shows up as a failing check rather than as a quietly worse CFG.

What is here

Ground truth lives in git; binaries do not. The JSON is small, reviewable, and belongs in history. The binaries are large and change a few times a year, so they ship as one tarball per configuration attached to a release. A configuration is an architecture and a compiler, <arch>-<compiler> (e.g., x64-gcc), and both ground-truth/ and binaries/ are split by it.

SHA256SUMS                        one line per binary; the integrity anchor
exclusions.json                   cases left out of scoring, and why
ground-truth/<config>/<case>.json one file per binary
schema/jumptable.schema.json      JSON Schema for the above
scripts/                          generate shasums, pack tarballs, find redundant cases

binaries/<config>/ is where a local build lands; it is gitignored.

Dataset Overview

Every case is a stripped PIE built at -O2, named <project>-<version>-<compiler>-pie-o2-<program>:

projects	GNU coreutils 9.9 and GNU binutils 2.45
compilers	gcc 16.2.0 and clang 23.0.0
architectures	x86 and x64, 139 binaries (119 from coreutils and 20 from binutils) each
tables	4,662 (x64) and 5,383 (x86)

• We exclude 43 coreutils programs (e.g., dir, vdir, [, sha1sum, true, yes) whose jump tables are already covered by another included program, to avoid redundant evaluation. Two tables count as the same when their kind, entry size, entry count, and targets relative to the indirect branch all match; scripts/find-redundant.py lists any case that adds no distinct table.